The jobs AI is creating

Buying an AI model is the easy part. Getting it to work with a hospital's records system, a bank's twenty-year-old database or a logistics firm's spreadsheets is not. A forward deployed engineer is sent to the customer to do that work. They sit with the customer's staff, connect the model to internal data and software, build the workflows people will actually use, and stay until it runs for real.

Palantir popularised the role. OpenAI, Anthropic, Cohere, Amazon Web Services and the finance company Ramp have since hired people under the same title. The job sits somewhere between engineer, consultant and product manager: part coder, part translator, part diplomat, and by week two, the person who knows where the client keeps the good coffee.

Large language models learn a lot from the internet, but not how a careful doctor weighs a diagnosis or how a tax lawyer reads a clause. To close that gap, AI labs pay people with real expertise to write example answers, grade the model's attempts and explain what a correct answer needs. The industry calls this reinforcement learning from human feedback.

At entry level it looks like labelling: tagging images or sorting text into categories. Further up, it's writing and judging long answers in one subject. At the top, doctors, lawyers and finance professionals review the model's work against the rules their fields run on. Most of it is project-based, remote and paid by the hour. Mercor, one of the firms that matches experts to labs, says it pays more than 30,000 contractors about $1.5 million a day.

Red teaming is an old security idea: hire people to attack your own defences. For AI, it means trying to get a model to do what it shouldn't. That could be giving out dangerous instructions, ignoring its safeguards, leaking data or doing harm while it runs as an agent. Then you write up exactly how it happened so the hole can be closed.

The UK's AI Security Institute, part of government, runs one of the best-known teams. It describes the work as finding and stress-testing "vulnerabilities in frontier AI systems" and sharing the results with AI companies and allied governments. Its teams cover jailbreaks, data poisoning and agent misuse, and test whether the monitors meant to catch bad behaviour can be beaten.

AI agents are programs that don't just answer questions but carry out multi-step tasks: processing an invoice, sorting sales leads, triaging a support ticket. Someone has to decide what each agent is allowed to do, check its work, handle the cases it can't and improve it week by week. In February 2026 the Harvard Business Review gave that person a name: the agent manager.

Drawing on Salesforce and other large companies, Suraj Srinivasan of Harvard Business School and Vivienne Wei of Salesforce describe agent managers as responsible for "orchestrating how AI agents learn, collaborate, perform, and work safely alongside humans". They compare it to the product manager, a job the software industry invented because it turned out to need one.

A few years ago, AI sat inside the IT or data team. Now many organisations have one senior person accountable for it. They choose which uses are worth paying for, set the rules, get the data ready and answer to the board when things go wrong.

In April 2025 the White House Office of Management and Budget told US federal agencies to name a chief AI officer by the end of June that year. In business, IBM's 2026 survey of 2,000 chief executives found that 76% of their organisations had one, up from 26% a year before.

Laws on AI are arriving faster than most companies can read them. The EU AI Act already bans some uses, and since August 2026 it has required that people be told when content is a deepfake or when they're dealing with an AI. Its heavier duties for high-risk systems, such as those used in hiring, lending or education, now start in December 2027 after a delay agreed this year. An AI governance lead keeps a list of every AI system the organisation uses, sorts them by risk and makes sure each one has the testing, paperwork and human oversight it needs.

Many come from privacy, where a similar job grew up around GDPR. The IAPP, the privacy profession's main body, added AI governance salaries to its survey for the first time in 2025, and it now runs a dedicated certification, the AIGP.

Plenty of organisations bought AI licences for everyone and then watched most people ignore them. An AI enablement lead runs the programme that changes that. They train staff, collect the uses that work into simple playbooks, run drop-in sessions and measure whether the tools save any time.

It's closer to training and development than to engineering. Job adverts in 2026 ask for people who can explain AI to non-technical teams and change habits across a whole organisation. The education software company Coursedog, for example, advertised for its first AI enablement lead to set out how AI would be used across the company.

Every AI model runs on racks of servers that need installing, cabling, cooling and repairing. Data centre technicians do that work. Around them, construction crews, electricians and commissioning engineers build the halls and test them before they open. It's shift work, on site, with tools.

The job isn't new, but AI has changed how much of it there is. LinkedIn's 2026 list of the fastest-growing US jobs included both data centre technicians and commissioning managers, who test new facilities such as data centres before they go live. Randstad looked at more than 50 million job postings and found that data centre construction workers earned an average of $81,800 a year, 32% more than other construction work.

In early 2024, an employee at the engineering firm Arup in Hong Kong joined a video call with what looked like the company's chief financial officer and several colleagues. Every one of them was a deepfake. The employee sent HK$200 million, about US$25 million, to the fraudsters. Cases like that have created work for people who can examine a piece of media and say, with evidence, whether AI made or altered it.

The work happens in fraud and security teams, newsrooms and fact-checking groups, platform trust and safety teams, and digital forensics for courts. Since 2 August 2026, the EU AI Act has required anyone using AI to make a deepfake to say so, which gives organisations that publish or host content a new reason to check.

Chatbots learned from the text on the internet. Robots have no internet of folded laundry and stacked boxes to learn from, so companies are recording people instead. A robot teleoperator wears motion-capture gear or a virtual reality headset and either does a task themselves or steers a robot through it, again and again. Every run becomes training data for the robot's AI.

In 2024 Tesla advertised for data collection operators for its Optimus humanoid robot, paying $25.25 to $48 an hour. The adverts asked for people between 5ft 7in and 5ft 11in who could walk for more than seven hours a day in a motion-capture suit and VR headset. The humanoid robot maker Figure lists jobs such as humanoid robot pilot and data creator on its careers board.

Legal AI tools can draft, summarise and search thousands of documents, but only a lawyer knows whether the result would survive a partner, a client or a judge. Legal engineers are qualified lawyers who work on the tools instead of the cases. They map how legal work really gets done, design AI workflows around it, run pilots with lawyers and fix what doesn't fit.

The legal AI company Harvey hires them. One of its 2026 adverts asks for a law degree or equivalent and at least three years in practice at a top firm or in-house, and offers on-target earnings of $220,000 to $320,000. Law firms are building the same skills inside. A&O Shearman, which Harvey says was the first law firm to roll out its tool across the firm, in December 2022, has a group of lawyers, software engineers and technologists working on generative AI.

For twenty years, marketers fought to get onto the first page of Google. Now more people ask ChatGPT, Gemini or Claude instead, and get one answer instead of ten blue links. An AI search specialist works on being part of that answer. They study which sources AI tools quote, make content clear enough for a model to pick up, and track how often the brand gets mentioned.

The field comes with awkward acronyms: GEO, for generative engine optimisation, and AEO, for answer engine optimisation. In February 2024 Gartner predicted that traditional search engine volume would fall 25% by 2026 as AI chatbots took over some searches. By mid-2026, one writer tracking job adverts had counted more than 50 of these roles at companies including Stripe, Amazon, Pfizer and HubSpot, with manager posts paying $100,000 to $174,000.

Businesses have insured against fire, flood and fraud for centuries. Insuring against an AI system making a costly mistake is new. An AI risk underwriter decides whether to cover a company's AI, on what terms and at what price. That means understanding how the model was tested, how often it fails and what happens when it does.

The reinsurer Munich Re wrote its first AI policy, for an anti-fraud model, in 2018. It now sells aiSure, which backs AI makers' promises about how their systems will perform, after technical checks. In April 2025, Armilla launched AI liability cover underwritten at Lloyd's of London by Chaucer, for AI that hallucinates, makes critical errors or doesn't perform as intended.

Hospitals and GP surgeries are bringing in AI scribes that listen to appointments and write up the notes. They save clinicians time, but a summary that leaves out an allergy is a patient safety problem. Clinical safety officers are qualified health professionals who assess those risks, sign off that a tool is safe to use locally, and keep watching once it's in use.

In England the role is written into the rules. NHS England's 2025 guidance on AI scribes says GP practices need a clinical safety officer to oversee clinical risk, and a formal risk assessment and safety case under the NHS standard DCB0160. It also says every output from an AI scribe must be reviewed by a clinician.

Every AI model runs on chips made in factories called fabs, where a single speck of dust can ruin a chip. Semiconductor technicians keep those factories running. They operate and maintain the machines, watch each step of production and catch faults before they spoil a batch. It's shift work in a full-body cleanroom suit.

The trade isn't new, but new chip factories and the demand for AI chips have made it hard to hire for. A 2023 study for the US Semiconductor Industry Association found that of about 115,000 new jobs expected in the US chip industry by 2030, around 67,000 were at risk of going unfilled, and 39% of that gap was technicians. Most of those roles need a two-year degree or certificate, not a university degree.

Video models can now generate shots, extend scenes and fill in backgrounds. Turning that into something worth watching still takes a director's eye, an editor's timing, and a lot of patience with a model that keeps giving the hero six fingers. Generative AI filmmakers combine traditional filmmaking with AI tools, and new studios are being built around them.

Asteria, a Los Angeles studio that calls itself "artist-led" and says its AI model is "clean and ethical", uses AI for jobs like in-betweening in animation, filling out crowds, extending shots and cleaning up footage. Promise, co-founded by the AI filmmaker Dave Clark, is aiming for films that Hollywood studios can release.

Every chatbot's tone is a design decision: how warm it is, how blunt, whether it pushes back when you're wrong. Get it wrong and the model becomes a flatterer that agrees with everything, which the industry calls sycophancy. A small number of people now work on this full time, writing the principles and example conversations that shape how a model behaves.

At OpenAI, a Model Behavior team of about 14 researchers shaped the personality of its models and worked on reducing sycophancy and political bias. In 2025 it was folded into the company's larger post-training group, a sign of how central the work had become. At Anthropic, the researcher Amanda Askell works on Claude's character, and has described the aim as something like a well-liked traveller who adapts to the people they meet "without pandering".

A few new schools split the day in two. Children spend about two hours on core subjects with AI tutoring software that works out what each child knows and moves at their pace. The rest of the day goes on workshops and projects. The adults are called guides, and according to Fortune they get to know and motivate students but don't plan lessons or grade homework.

The best-known example is Alpha School in the US, which Fortune reports teaches more than 1,200 students and charges up to $75,000 a year. Its founder, MacKenzie Price, told CBS News that guides earn six-figure salaries. The model has critics: a Northwestern University researcher told CBS that at these prices "it's not going to be available to everybody".